On Tue, Jul 07, 2020 at 11:01:38PM -0400, Liang, Kan wrote:
> On 7/7/2020 3:48 PM, Bjorn Helgaas wrote:
> > [+cc Stephane in case he has thoughts on the perf driver claim issue]
> >
> > On Thu, Jul 02, 2020 at 10:05:11AM -0700, kan.liang@xxxxxxxxxxxxxxx wrote:
> > > From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > >
> > > On Snow Ridge server, several performance monitoring counters are added
> > > in the Root Port Configuration Space of CPU Complex PCIe Root Ports A,
> > > which can be used to collect the performance data between the PCIe
> > > devices and the components (in M2IOSF) which are responsible for
> > > translating and managing the requests to/from the device. The
> > > performance data is very useful for analyzing the performance of the
> > > PCIe devices.
> > >
> > > However, the perf uncore driver cannot be loaded to register a
> > > performance monitoring unit (PMU) for the counters, because the PCIe
> > > Root Ports device already has a bonded driver portdrv_pci.
> > >
> > > To enable the uncore PMU support for these counters on the uncore
> > > driver, a new solution should be introduced, which has to meet the
> > > requirements as below:
> > > - must have a reliable way to find the PCIe Root Port device from the
> > > uncore driver;
> > > - must be able to access the uncore counters of the PCIe Root Port
> > > device from the uncore driver;
> > > - must support hotplug. When the PCIe Root Port device is removed, the
> > > uncore driver has to be notified and unregisters the uncore PMU.
> > >
> > > A new platform device 'perf_uncore_pcieport' is introduced as part of
> > > the new solution, which can facilitate the enabling of the uncore PMU in
> > > the uncore driver. The new platform device
> > > - is a child device of the PCIe Root Port device. It's allocated when
> > > the PCIe Root Ports A device is probed. (For SNR, the PMU counters are
> > > only located in the configuration space of the PCIe Root Ports A.)
> > > - stores its pdev as the private driver data pointer of the PCIe Root
> > > Ports A. The pdev can be easily retrieved to check the existence of
> > > the platform device when removing the PCIe Root Ports A.
> > > - is unregistered when the PCIe Root Port A is removed. The remove()
> > > method which is provided in the uncore driver will be invoked. The
> > > uncore PMU will be unregistered as well.
> > > - doesn't share any memory and IRQ resources. The uncore driver will
> > > only touch the PMU counters in the configuration space of the PCIe
> > > Root Port A.
> >
> > I have to admit this is clever. I don't really *like* it, but we
> > don't have any very good alternatives at the moment.
> >
> > I don't like the idea of a list of PCI IDs
> > (perf_uncore_pcieport_ids[]) below that must be updated for every
> > device that needs something like this. That PCI ID information is
> > normally in the drivers themselves, not in bus-level code like this.
>
> I don't want to create a platform device for every single device. So I added
> a check here. Yes, it doesn't look pretty, but I don't have a better
> solution for now.
I do not want to merge a stream of device IDs in portdrv_pci.c. As
far as I'm concerned, that's a non-starter. New devices should not
require changes in the PCI core.
> > And I don't like the way this subverts the device ownership model.
> > Now we have several drivers (pciehp, aer, dpc, etc, plus this new perf
> > driver) that share the same PCI device. And we rely on the assumption
> > that none of these drivers interferes with the others.
> >
> > I think the best way to deal with this would be to incorporate the
> > existing portdrv users (pciehp, aer, dpc, etc) directly into the PCI
> > core so portdrv would not use pci_register_driver(), leaving the Root
> > Port device available for the perf driver to claim it the normal way.
> > But realistically I don't know when or even whether this will be done.
> >
> > I think Stephane has worked around this problem in a different way,
> > IIRC by using pci_get_device() in a perf driver to find Ports of
> > interest. That also subverts the device ownership model, and it
> > doesn't work naturally with hotplug, but at least it gets the device
> > IDs out of the PCI core and into the driver where they belong. And
> > there's value in solving the same problem in the same way.
> >
> > Wait a minute! You've already used the pci_get_device() strategy
> > several times:
> >
> > 2b3b76b5ec67 ("perf/x86/intel/uncore: Add Ice Lake server uncore support")
> > fdb64822443e ("perf/x86: Add Intel Tiger Lake uncore support")
> > ee49532b38dd ("perf/x86/intel/uncore: Add IMC uncore support for Snow Ridge")
> >
> > So what's really different about *this* situation? Why would you not
> > just continue using the same strategy?
>
> There are three different methods (MSR, PCICFG, and MMIO) to access the
> uncore counter, while each counter can only accessed using one of the
> method. Many devices have uncore counters. All counters on the same device
> must be accessed using the same method.
>
> The perf uncore driver abstracts three code paths for the above three
> different methods. Each code path is shared among the devices which have the
> corresponding access method.
>
> The pci_get_device() strategy is used by the device in which counters can be
> accessed by MMIO. Currently, the only such device is the IMC (Integrated
> Memory Controller) device. The problem is that the BAR address of the IMC
> counters is located in the PCI Configuration Space of the Configuration
> Agent (Ubox) device. The perf driver is not supposed to bind the Ubox device
> while accessing the counters in another device, the IMC device. So the
> pci_get_device() is used to retrieve the pci_dev of the Ubox. The perf
> driver reads the BAR address from the Ubox and maps it to access the IMC
> counters.
>
> The counters in the Root Port device are in the PCI configuration space of
> the device. For a device whose counters are in the PCI configuration space
> of itself, the perf driver should probe and bind the device. However, the
> Root Port device is already bound by the portdrv driver.
Understood. That's why I said above that the current portdrv design
is a problem.
> To maximize the code reuse in the perf driver, a platform device is
> introduced as a child of the Root Port device. So the perf driver can probe
> and bind the platform device in a similar method.
The current portdrv design has two struct devices for each port: the
usual one in struct pci_dev and a second in struct pcie_device.
That's already one too many, and this platform device plan would
add a third one.
It also would add an arbitrary new way to bind a driver to a PCI
device, so we end up with *two* drivers bound to the same device. I
don't think that's a good design.
> The pci_get_device() strategy may work for the case, but from the
> perspective of the perf driver, I don't think it's a good solution. We have
> to specially handle the pci_get_device() strategy in the code path of the
> PCICFG access method. Also, it's not a complete solution, e.g. as you said
> it doesn't work naturally with hotplug.
I think pci_get_device() is also a pretty poor solution, but I think
it's the best of the current poor options because it's already pretty
widely used by EDAC drivers and others, and at least it doesn't give
the illusion that we're using a safe driver binding model.
I think the hotplug issue could be handled with bus-level
notifications (also ugly, but at least workable). xen_pci_notifier()
is a possible example.
I would welcome any help in reworking portdrv so we could do this more
cleanly.
> > > Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > > ---
> > > drivers/pci/pcie/portdrv_pci.c | 38 ++++++++++++++++++++++++++++++++++++++
> > > 1 file changed, 38 insertions(+)
> > >
> > > diff --git a/drivers/pci/pcie/portdrv_pci.c b/drivers/pci/pcie/portdrv_pci.c
> > > index 3acf151..47e33b2 100644
> > > --- a/drivers/pci/pcie/portdrv_pci.c
> > > +++ b/drivers/pci/pcie/portdrv_pci.c
> > > @@ -15,6 +15,7 @@
> > > #include <linux/init.h>
> > > #include <linux/aer.h>
> > > #include <linux/dmi.h>
> > > +#include <linux/platform_device.h>
> > > #include "../pci.h"
> > > #include "portdrv.h"
> > > @@ -90,6 +91,40 @@ static const struct dev_pm_ops pcie_portdrv_pm_ops = {
> > > #define PCIE_PORTDRV_PM_OPS NULL
> > > #endif /* !PM */
> > > +static const struct pci_device_id perf_uncore_pcieport_ids[] = {
> > > + { PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x334a) },
> > > + { },
> > > +};
> > > +
> > > +static void perf_platform_device_register(struct pci_dev *dev)
> > > +{
> > > + struct platform_device *pdev;
> > > +
> > > + if (!pci_match_id(perf_uncore_pcieport_ids, dev))
> > > + return;
> > > +
> > > + pdev = platform_device_alloc("perf_uncore_pcieport", PLATFORM_DEVID_AUTO);
> > > + if (!pdev)
> > > + return;
> > > +
> > > + pdev->dev.parent = &dev->dev;
> > > +
> > > + if (platform_device_add(pdev)) {
> > > + platform_device_put(pdev);
> > > + return;
> > > + }
> > > +
> > > + pci_set_drvdata(dev, pdev);
> > > +}
> > > +
> > > +static void perf_platform_device_unregister(struct pci_dev *dev)
> > > +{
> > > + struct platform_device *pdev = pci_get_drvdata(dev);
> > > +
> > > + if (pdev)
> > > + platform_device_unregister(pdev);
> > > +}
> > > +
> > > /*
> > > * pcie_portdrv_probe - Probe PCI-Express port devices
> > > * @dev: PCI-Express port device being probed
> > > @@ -113,6 +148,8 @@ static int pcie_portdrv_probe(struct pci_dev *dev,
> > > if (status)
> > > return status;
> > > + perf_platform_device_register(dev);
> > > +
> > > pci_save_state(dev);
> > > dev_pm_set_driver_flags(&dev->dev, DPM_FLAG_NO_DIRECT_COMPLETE |
> > > @@ -142,6 +179,7 @@ static void pcie_portdrv_remove(struct pci_dev *dev)
> > > pm_runtime_dont_use_autosuspend(&dev->dev);
> > > }
> > > + perf_platform_device_unregister(dev);
> > > pcie_port_device_remove(dev);
> > > }
> > > --
> > > 2.7.4
> > >
